Role-scoped AI assistant
Five specialists in one chat
Per-specialist skill gating means an assistant cannot edit data outside its role. Forty-seven tool-bound skills, zero cross-domain hallucinations.
Read the caseAll case studies
Case study · Intake to tracker
Meridian Operations · multi-team change intakeVague request to filed work item: 5 minutes
A stakeholder opens the assistant, picks the Business Analyst, answers a structured playbook, and gets a filed item with a display key and a handoff offer before the coffee cools.
0 min
Vague request to filed work item
0%
Deduplication check before filing
0.00
Acceptance-criteria coverage on dev rubric
Built with
The problem
A one-line message at 9:12am, seven replies by 9:45, engineering gets it Friday
The same request shape kept happening. A stakeholder writes 'can we add X' in a shared channel. Someone asks 'can you write up a spec.' The stakeholder writes a paragraph that misses the acceptance criteria. Three rounds of clarifying questions later, the item reaches the tracker — usually two weeks later, sometimes never. The product lead spent Monday mornings chasing what was missing.
- The one-liner problem'Can we make the search faster' arrived in the shared channel at 9:12am. By 9:45am the thread had seven replies and no one had written down what 'faster' meant, who was affected, or whether this was a bug or a feature. Engineering got it Friday.
- Intake fatigueThe product lead spent two hours every Monday morning chasing missing information on items filed the previous week. The same questions — 'what's the current behavior,' 'who does this affect,' 'what would success look like' — asked on repeat, per item, per team.
- Spec driftA stakeholder filed a change request, the item sat for two weeks, and by the time engineering started, the underlying behavior had changed. The spec was out of date. The implementation created a regression in an adjacent component.
- Duplicate filingsTwo teams filed the same request under different titles three times in one quarter. The duplicates weren't caught until sprint planning, when three nearly identical items appeared in the same column and the team had to triage backwards.
#shared-channel · 33-minute thread
no item filed@alex9:12can we make search faster
@jamie9:14faster how — slower than 5s? 10s?
@alex9:18idk, just slow
@sam9:21is this a bug or a feature ask?
@morgan9:30we logged something similar last sprint
@alex9:42different page i think
@jamie9:45ok someone file a ticket
33 minutes · 7 messages · 0 acceptance criteriaEng gets it Friday
The pipeline
From inbox to verified record in one pass
Six steps from picker to filed item. The interview playbook is type-specific; the deduplication check runs against the live index.
01Pick
Business Analyst specialist selected
Stakeholder opens the assistant and picks the BA specialist. The first message is a type-classification question, not a blank prompt.
02Classify
Type-specific playbook loaded
Specialist classifies the ask into bug, feature, enhancement, or change request. Each type loads a different interview playbook.
Claude
03Interview
Type-specific questions, one at a time
The playbook walks the stakeholder through what the dev team needs: who is affected, current vs. desired behavior, blast radius, success criteria, out-of-scope, risks. Specialist does not move on until each answer is non-empty.
Claude
04Emit
Structured spec drafted
Specialist emits a user story, three to seven When/Then clauses, at least one test case per clause, and a doc-update list naming the specific files that will need post-change maintenance.
Claude
05Deduplicate
Tracker searched against the live index
Before filing, the specialist searches the live tracker for matches above a similarity threshold. Surfaces any matches and asks: merge or proceed as new?
PostgreSQL
06File
Work item filed; handoff offered
Tracker assigns a display key (e.g., IQUE-142). Specialist confirms with the key and offers an immediate handoff to the implementation orchestrator.
Microsoft Teams
Complete answers
Filed in under five minutes with full rubric coverage
Stakeholder provides complete answers. Specialist emits a draft, deduplication finds no conflicts, and the item is filed in under five minutes with all rubric fields populated.
Incomplete answers or dedup match
Pause and explain — never file an ambiguous item
If the stakeholder gives incomplete answers or the deduplication step surfaces a potential match, the specialist pauses, explains the conflict or the gap, and waits. No item is filed until the stakeholder resolves the ambiguity.
Validation review
Per-element confidence on the emitted spec
The spec has five elements. The doc-update list is the lowest-confidence one — and the source explains exactly why.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
Type classificationbug · feature · enhancement · change
97%High
User storyAs X I want Y so that Z
95%High
When/Then clausesConditional acceptance criteria
91%High
Test casesVerifiable per clause
88%High
Doc-update listSpecific files that will need maintenance
62%Low
Routed to human review. Naming specific files requires reading the actual repo. The specialist has a snapshot and sometimes names files that have moved or been renamed since the snapshot was taken. Verify doc-update lists before closing an item.
4 of 5 fields cleared the 0.85 threshold
model: Spec-element confidenceThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Claude
Classifies, interviews, drafts the spec, and reasons about deduplication
PostgreSQL
Work-item table, display-key counter, tracker search index
Drizzle ORM
Type-safe queries against the work-item and search tables
Microsoft Teams
Handoff notification when a new item is filed and orchestrator is offered
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed when the bottleneck moved from intake to capacity
Same-week implementation rate climbed because items arrived at engineering complete. The bottleneck shifted from intake to engineering capacity.
0 min
Vague request to filed item
0%
Items deduplicated before filing
0.00
Acceptance-criteria coverage on dev rubric
0%
Items implemented same-week (vs. multi-week before)
Watch acceptance-criteria coverage on the dev rubric
Currently 0.92 — roughly one in twelve items reaches engineering with at least one rubric field missing. Target is 0.95. Watch the clarification-question rate from engineering as the leading indicator: when devs start asking 'what does success look like' or 'who is affected,' the playbook needs new examples drawn from the questions that slipped through.